A SEGMENTED LINEAR SUBSPACE MODEL FOR ILLUMINATION-ROBUST FACE RECOGNITION

摘要

Compensating for illumination variations has crucial importance for a face recognition system because it is hard to make any assumptions about illumination in realistic environments where face recognition systems would be deployed. In this paper, we describe a Segmented Linear Subspace model for illumination-robust face recognition where we model the images of a face using a collection of linear subspaces. This model is a generalization of the low-dimensional linear subspace models, and is motivated by the structure of the illumination process that introduces higher correlation into the pixels that have similar surface normals. We propose empirical procedures to determine the optimal number of dimensions for the linear subspaces and the number of regions in the segmentation to obtain the best performance. We perform extensive experiments to demonstrate that this model provides a simple and powerful method for illumination-robust face recognition.